Icon

LLM Fraud_​Detection - Embeddings

Outlier Detection: Embedding BasedOutlier detection as a proxy for potential fraud detection. We try to detect numerical outliers in the high-dimensional embedding space.In the top right of this section we rank the emails numerically (in the multidimensional embedding space), based on how much they diverge from the rest of the emails; this can allow us to notice potentially suspicious activity. The lower two blocks allow us to visualise this information, in either two or three dimensions. Note that we apply PCA dimension reduction, which basically means we squeeze down the information, from 1500 dimensions, to 2/3 dimensions. Not all of the information is retained in the final plot, but it still allows us to detect outliers to a reasonable degree.

Select emails

Right click -> Interactive View,
Select some rows then click 'Apply'

PCA Dimension Reduction and Plotting

Project all embeddings on to common space so as to remove influence of components which we are not interested in.
This method is efficient computationally.

2D
(Right click on 'Scatter Plot'->'Interactive View...')

3D
(Right click on '3D Scatter'->'Interactive View...')

Cluster Similar Items Together

You can inspect the cluster it belongs to in the output

Additionally you can configure the clustering by changing the settings.

Set up LLM

Credentials and setup

Load Email Dataset

String Manipulation
Take onlyuser-selectedrows
Rule-based Row Filter
User canselect whichemails toanalyse
Table View (JavaScript)
Reduce to 2-dimensions
PCA Apply
Scatter Plot (JavaScript)
Missing Value
assign colors to classes
Color Manager
Reduce to 3-dimensions
PCA Apply
Fitting/ 'Preprocess'for PCA
PCA Compute
OpenAI Authenticator
CSV Reader
Fetchembeddings
Text Embedder
OpenAI Embeddings Connector
Constant Value Column
3D Scatter Plot (Plotly)
Split Collection Column
assign colors to classes
Color Manager
Credentials Widget
Column Filter
Column Filter
k-Means

Nodes

Extensions

Links